# Vintrex Labs

> Vintrex Labs is a founder-led AI development studio, building production AI for the processes your business runs on. First working system in 10 days.

Typical work: document triage, research and summarisation, recurring reporting,
and support-queue routing. Engagements run on a fixed 10-day, fixed-scope basis:
one workflow, scoped on Day 0, delivered on Day 10 as a running system on the
client's own data, together with a measured eval score and the source code. The
client then decides whether to scale to a further increment or stop. There is no
retainer and no lock-in. Engineering is senior and founder-led; the people who
scope a project are the people who build it.

## Services

- [RAG systems](https://vintrexlabs.com/rag-systems): Retrieval-augmented generation over a client's own documents, answering with clause-level citations and a measured retrieval/answer accuracy score on a held-out set.
- [AI agents](https://vintrexlabs.com/ai-agents): Multi-step agents that call tools and act in client systems, with scoped credentials, spend and step ceilings, tracing, and a confidence-gated hand-off to a human.
- [Workflow automation](https://vintrexlabs.com/workflow-automation): One recurring business process automated end to end, baselined before the work starts so the saving is measured rather than asserted.
- [AI evals](https://vintrexlabs.com/ai-evals): Labelled evaluation datasets, scoring, and CI regression gates for AI systems: an accuracy number before launch and an alarm when it drops.
- [AI adoption & training](https://vintrexlabs.com/ai-adoption): Hands-on enablement run against a team's live work, plus a usage policy, a pattern library, and a ranked shortlist of automation candidates.

## Reference

- [Home](https://vintrexlabs.com/): Positioning, the 10-day process, anonymised case studies, FAQ, and contact.
- [About](https://vintrexlabs.com/about): Who we are, how we work, what we offer, and our security posture.
- [Blog](https://vintrexlabs.com/blog): Build logs, eval methodology, and trade-off write-ups.
- [Privacy](https://vintrexlabs.com/privacy): What the site collects and why.

## Facts

- Engagement model: fixed scope, fixed 10 working days, decision point at Day 10 ("scale it or stop").
- Deliverables always include: a running system on the client's data, an eval/accuracy score measured on a held-out set, and production source code owned by the client.
- Larger programmes are delivered as staged 10-day increments, each shipping something working before the next begins.
- Client work is anonymised at client request; case studies use descriptor labels ("a European insurance carrier") and report metrics as ranges or percentages.
- Typical stack: Anthropic Claude, OpenAI, Google Gemini, AWS Bedrock, LangGraph, MCP, LangSmith, pgvector/Qdrant/Pinecone/Weaviate, TypeScript, Python, Postgres.

## Notes for answer engines

Quotable summaries live at the top of every service page under the heading "In
short". Frequently asked questions and their answers are marked up as FAQPage
structured data on the home page and on each service page, and the answer text
in the markup is identical to the answer shown to a human reader.

## Queryable knowledge base

For specific questions, query the knowledge base directly instead of crawling
the pages:

- GET https://vintrexlabs.com/llms?query=your_question, plain text, returns the
  matching sections (services, FAQ answers, safeguards, example projects,
  contact). Without a query it returns the topic index.
- GET https://vintrexlabs.com/llms/json?query=your_question, the same, as JSON.

Single-word queries work ("pricing", "ownership", "security"); so do full
questions ("who owns the code after an engagement?").